Assuming we have 1296 images (36x36) around the object in a 360x360 setup, or perhaps slightly fewer, say 100 images.
The current algorithm generates pairs for all possible combinations, resulting in a substantial number of pairs (100*99).
Without a proper sampling policy, this can be challenging to handle.
A small recommendation would be to initiate and maintain a pair-loss table from scratch, gradually increasing the number of pairs and sampling based on the convergence of loss.
Assuming we have 1296 images (36x36) around the object in a 360x360 setup, or perhaps slightly fewer, say 100 images. The current algorithm generates pairs for all possible combinations, resulting in a substantial number of pairs (100*99). Without a proper sampling policy, this can be challenging to handle. A small recommendation would be to initiate and maintain a pair-loss table from scratch, gradually increasing the number of pairs and sampling based on the convergence of loss.